← All models

DeepSeek-V3.2

Published Publisher DeepSeekdeepseek-chat

Overview

DeepSeek-V3.2 is a general-purpose, reasoning-oriented language model designed to balance reasoning ability and output length for everyday questions and agent tasks. It was released on December 1, 2025. (deepseek.com)

Capabilities and uses

  • Supports thinking and non-thinking modes, including tool use in both modes. Suitable applications include question answering, coding and tool-using agents. (api-docs.deepseek.com)
  • Uses DeepSeek Sparse Attention to improve efficiency for long-context workloads. The official model card specifies text input and output, a 128K context length, 671 billion total parameters and 37 billion activated parameters per token. (fe-static.deepseek.com)
  • Model weights and code are available under the MIT License for self-deployment. (fe-static.deepseek.com)

API status and limitations

At release, deepseek-chat selected V3.2's non-thinking mode; deepseek-reasoner selected its thinking mode. On April 24, 2026, DeepSeek said those names had been redirected to V4-Flash and were scheduled for discontinuation on July 24, 2026. Therefore, deepseek-chat should not be assumed to provide V3.2 today. (api-docs.deepseek.com)

The model is text-only. Outputs can contain errors and should be checked, especially for consequential decisions. DeepSeek also notes that its example output parser requires robust error handling for production use. (fe-static.deepseek.com)

Pricing

Checked November 17, 2026: Current official V3.2 input and output API prices could not be verified in either CNY or USD. The current Chinese and international pricing pages list newer V4 models, so their prices should not be attributed to V3.2. (api-docs.deepseek.com)

Official sources